Skip to content
All posts
2 min read

Spec-Driven Development Is How You Get Value From AI Coding

AI-assisted development has been in my production workflow since 2022. The engineers getting 10x from it aren't prompting harder — they're specifying better.

I've used AI-assisted development in production since 2022 — first GitHub Copilot on enterprise codebases, now agentic tools that write whole modules. The pattern that separates teams getting real leverage from teams getting slop is not prompt cleverness. It's whether a spec exists before the code does.

Why specs and AI are a natural pair

An LLM is a powerful but literal collaborator: it does exactly what the context implies, and fills every gap with a guess. A written spec removes the gaps. When I started my quant trading platform, the first commit wasn't code — it was SYSTEM_DESIGN.md: module boundaries, data contracts, configuration strategy. Every AI-assisted session after that had a source of truth to be checked against.

The workflow that works for me:

  1. Write the contract — what the module does, its inputs/outputs, its failure modes. Plain markdown, versioned with the code.
  2. Direct, don't dictate — give the AI the spec and the relevant code, ask for an implementation, and let it handle the mechanical breadth.
  3. Verify like a reviewer, not a spectator — run it, test the edges the spec calls out, and reject anything that satisfies the letter of the prompt but not the contract.

The spec is the durable artifact

Here's the inversion that took me a while to see: with AI in the loop, code is becoming cheap and specs are becoming the asset. Code can be regenerated from a good spec in minutes. A spec can't be regenerated from code without losing the why — the rejected alternatives, the constraints, the intent.

This also explains why senior engineers benefit from AI more than juniors, which is the opposite of what everyone predicted. The bottleneck was never typing speed; it was knowing precisely what to build. Nine years of writing API contracts, reviewing designs, and debugging production incidents is exactly the training you need to direct an AI well.

What I don't delegate

Three things stay human in my workflow: architecture decisions (the AI proposes, the spec decides), security-sensitive code paths (auth, payments — reviewed line by line, always), and the definition of done (tests the AI writes to satisfy its own code don't count as verification).

AI-assisted engineering isn't a junior engineer that never sleeps. It's a force multiplier on clarity. If your thinking is vague, it multiplies the vagueness — and ships it.